I am working on a stock price prediction project where I am trying to use an LSTM to predict stock prices. Currently, I am using trying to use 10 days of data to predict the 11th day’s stock price. To generalize my question regarding the target variable, let’s say I have the following

[1, 2, 3, 4, 5]
[2, 3, 4, 5, 6]
...

as my feature data where each list shown is a datapoint with 5 timesteps. Should my target variable be

[6]
[7]
...

or

[2, 3, 4, 5, 6]
[3, 4, 5, 6, 7]

I’ve tried both and have gotten better results with the second, but I am unclear what is the best approach. In regards to my question about return_sequences, I understand that when return_sequences=False, the last timestep is outputted for each data point. As a result, if return_sequences=True, the last timestep for each output should be equal to the output when return_sequences=False. However, I am getting different results.

Below is my code:

keras.utils.set_random_seed(420)

num_features = len(features)

callback = keras.callbacks.EarlyStopping(monitor='val_loss',
                                         patience=10,
                                         restore_best_weights=True)

model = Sequential()
model.add(LSTM(units=128))
model.add(Dense(units=1))
model.compile(loss='mean_squared_error', optimizer='adam')
history = model.fit(X_train_final, y_train_final, 
                    epochs=50, 
                    batch_size=64, 
                    verbose=2, 
                    validation_data = [X_test_final, y_test_final],
                    callbacks=[callback])

test_predict = model.predict(X_test_final)

I then use

test_predict = test_predict[:,-1,:]

when return_sequences=True otherwise I leave it out if return_sequences=False. I get different results for each.

Any help would be much appreciated!